[chaoming0625]

Fixes #187.

Summary

• _model_that_not_allow_param_assign (_compiler/module_info.py) now lists every ParamState written inside the traced step by its model path. A state not reachable from the model falls back to its repr. The message says the rule applies whether or not the parameter is trained online, and gives the two supported patterns:
  • trained weights: update them outside the step, from the gradients etrace_grad returns;
  • locally-updated (plastic) weights: hold them in a non-param state such as brainstate.LongTermState, which is then not trained online.
• docs/advanced/limitations.ipynb: new section "Parameters Are Constant Within a Step" (failing + working example, stored output) and a matching Summary item.
• Spec: docs/specs/2026-10-24-issue-187-param-write-message.md.

Behaviour is unchanged: the same NotSupportedError is raised in the same cases.

Before / after (issue repro)

# main
w_local as ParamState   : NotSupportedError: The parameter state "ParamState(f32[3,4])" is rewritten in the model. Currently, the online learning method we provided does not support the dynamical weight parameters.
# this branch
w_local as ParamState   : NotSupportedError: ParamState ('w_local',) is written inside the online step. braintrace requires parameters to be constant within a step (the eligibility-trace recursion assumes it), whether or not they are trained online. Update trained weights outside the step, from the gradients returned by `etrace_grad` (e.g. with an optimizer), or hold a locally-updated weight (e.g. a plasticity rule) in a non-param state such as `brainstate.LongTermState`; it is then not trained online.
w_local as LongTermState: compiled

Tests

module_info_test.py::TestParamStateWriteError:

• the message names the path and both patterns;
• all written params are listed;
• nested sub-module path ('layer', 'w_local');
• unattached global falls back to repr;
• the LongTermState variant compiles.

Full suite (CPU, remote): 2911 passed, 1 skipped. main gives 2906 passed. The only collection error is a missing hypothesis in the remote env, and it also occurs on main.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!